moonIsHere007/smai-a4

SMAI Assignment 4, Q1: MLP and IOC-NN on CIFAR-10

File	Model	Test accuracy
torch_mlp_p0.npz / .json	PyTorch MLP 3072-256-10, ReLU, trained 100 epochs on true labels	0.5086
torch_ioc_p0.npz / .json	PyTorch IOC-NN 3072-256-10, ELU, weights of layer 2 ≥ 0	0.4998
torch_preprocessing.npz	column mean / std of the 50,000 training images (for both models above)
mlp_numpy_selected.npz / .json	NumPy MLP selected in 1.2.5 (M2, 3072-256-10, ReLU)	0.5048
mlp_numpy_selected_preprocessing.npz	column mean / std of the 36,000-image train split

Every model has 789,258 parameters. The .npz files contain W0, b0, W1, b1 with W of shape (n_in, n_out), so the logits are relu_or_elu(x @ W0 + b0) @ W1 + b1. The input x is the 32×32×3 image flattened to 3072 values (HWC order), divided by 255 and standardised with the stored mean / std. The JSON files describe the architecture, the preprocessing and the training settings (seed 1101).